




版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领
文档简介
1、The Design of Desired Collectives with Multi-Agent Simulation,Akira Namatame Dept. of Computer Science National Defense Academy, Japan namanda.ac.jp,2,Collectives of Interacting Agents,Collective of interacting agents is complex with the following properties: (1) Non-linearity and path-dependency (2
2、) Self-organization (3) Emergence (4) Unintended consequence,We propose the approach of designing desired collectives with the agent-based simulation,3,preference,interest,goal,Agent,Collective,Agents Behavior Based on the Logic of Minority, Agents gain if they take the same action as minority does.
3、,(1) Purposive decision Decision based on preference or interest (2) Contingent decision Decision based on what others are doing,4,Highlights of The Talk, Characterize the inverse and forward problems to self-organize desired collectives Propose the interactive design with multi-agent simulation.,5,
4、Logic of Minority: Symmetric Problem (1),At each time step,agents make a binary choice : Agents on the minority side get more payoffs than those who are the majority side.,Minority games El Farol bar problem,U(S1)=a(1-n/N) U(S2)=b(n/N),6,Logic of Minority: Asymmetric Problems,Congestion problem,Mark
5、et entry games,Market,S1 : use a car S2 : use a train,payoff=benefit - time,7,Reasons for Undesirable Outcomes, (1) Bounded rationality of agents (2) Inconsistency between individual rationality and group rationality (1) Agents behave with false rules How do agents learn desirable rules? (2) Agents
6、behave with inappropriate utility functions. How do agent should modify their own utility functions?,8,Symmetric Problem vs. Asymmetric Problem,U(S1)= U(S2),(1) Nash equilibrium:,(2) Pareto optimal:,Average utility E=pU(S1)+(1-p)U(S2) =(a+b)(p-p2) Average utility is maximized at p=0.5,Average utilit
7、y is maximized,Average utility E=pU(S1)+(1-p)U(S2) =a(p-p2)+b Average utility is maximized at p=a/2(a+b),9,Decomposition to Pair-wise Problems,U(S1)= a(1-n/N) U(S2)= b(n/N),(1) Symmetric problem,(2)Asymmetric problem,U(S1)= a(1-n/N) U(S2)= b,q=a/(a+b),1-q,10,Desirable Collective: Stability, Efficien
8、cy, Fairness,Stability: Desirable collective need to be equilibrium of underlying games Efficiency: Desirable collective need to be efficient of underlying games Fairness Since there are many equilibria, the criteria of stability and efficiency are not enough, and fairness is evolutions solution to
9、the equilibrium selection problem,11,Characterization of Learning Models,(1) Learning models without coupling with others Reinforcement learning Agents reinforce the strategy which gains the payoff Evolutionary learning Agents evolve strategy of interaction (2) Learning models with coupling Best-res
10、ponse learning Agents adapt based on the best-response strategy,12,Agents Make Choices without Coupling,There is no coupling,agent has several randomly generated strategies of memory m. A each step, the player uses the strategy that would have maximized its gains over the entire history.,Most common
11、 learning model in minority games,13,Coupling of Agents,(1) Coupling with collectives (2) Coupling with neighbors,14,Coupling Rule between Two,Agents make choice based on the past two history,Coupling rule between agents,15,The Performances of Evolutional Learning,Noise=0%,Noise=5%,Max,Min,Ave,Max,M
12、in,Ave,16,What Agents Acquired with Evolutinary Learning ?,400 agents with different rules at the beginning evolved to share one of 15 coupling rules.,The number of agents,17,Commonality of Acquired Rules,The 15 meta-rules shared by all agents have the commonality,18,Coupling with Local Neighbors,S1
13、,: The proportion of neighbors to choose,The behavioral rule as give-and-tale,19,Simulation Results,Efficient and equitable dynamic orders are emerged with give-and-take,S1,S2,20,Coupling Agents with Collectives,(1)The action variable of agent Ai , a1(t) = 1 : S1 (Go) a1(t) = 0 : S2 (Stay) (2)The St
14、atus of the Bar,The bar is crowded at time t,The bar is not crowded at time t,(3) Rules of give-and-take,If gain, then yields, if no gain, chooses randomly,21,Simulation Results (=0.5),Blue line;S1, Red line;S2,All agents choose Nash strategies,Payoff distribution,Give & Take Learning,22,Efficient U
15、tilization of Limited Resource with Too Many Contestants,Market entry games El Farol bar problem,The capacity of resource: q The capacity of resource: q/2,How limited resource is maximally utilized under an efficient and equitable situation?,Payoff,23,How to Solve Inverse Problem?,(1)Design right be
16、havioral rules Interacting agents need to develop right behavioral rules for desirable collectives (2) Design right utility functions Agents need to modify their endogenous utility functions for desirable collectives.,24,Exogenous Design with Subsidy or Tax,How should utility functions be redesigned
17、 with subsidy or tax?,Payoff,U(S1)=1-n/N U(S2)=1-q,U(S1)=1-n/N (n/N)q/(2-q),(n/N)q/(2-q): Tax,Nash equilibrium: n/N=q Pareto-optimal: n/N= q/2,U(S2)=1-q + q/2,q/2: Subsidy,25,Endogenous Design with Give&Take,The capacity of resource (bar) : Nq,(Case 2) Agents who chose S1(enter), choose S2(stay) A p
18、art of agents who chose S2(stay) choose S2(stay) again,The number of agent who stayed at time t,(Case 1) Agents who chose S2(Stay), choose S1(Enter) A part of agents who chose S1(Enter), choose S1(Enter) again.,26,Solving Inverse Problems with Agent-based Simulation (ABS),Evolutionary Design with Agent-Based Simulation,27,Conclusion: Achieving Desired Collectives,We showed that collective behavior with the logic of minority is much
温馨提示
- 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
- 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
- 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
- 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
- 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
- 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
- 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。
最新文档
- 防沙治沙工程投标书(参考)
- 风力机叶片设计-洞察及研究
- 2025年医学高级职称-放射医学(医学高级)历年参考题库含答案解析(5卷单项选择题100题)
- 2025年医学高级职称-医院药学(医学高级)历年参考题库含答案解析(5卷100题)
- 2025年医学高级职称-中西医结合内科(医学高级)历年参考题库含答案解析(5卷单选100题)
- 2025年住院医师规范培训(各省)-重庆住院医师内分泌科历年参考题库含答案解析(5卷单选100题)
- 2025年住院医师规范培训(各省)-甘肃住院医师呼吸内科历年参考题库含答案解析(5卷单选100题)
- 2025年住院医师规范培训(各省)-江苏住院医师神经内科历年参考题库含答案解析(5卷单选一百题)
- 2025年住院医师规范培训(各省)-江苏住院医师中医内科历年参考题库含答案解析(5卷单选一百题)
- 2025-2030全挂车市场前景分析及投资策略与风险管理研究报告
- 二维材料在柔性电子中的应用研究
- 内科患者VTE风险评估表
- 一年级上册美术教案-第1课 让大家认识我:诚实最好 ▏人美版
- 科学认识天气智慧树知到期末考试答案2024年
- (高清版)DZT 0064.15-2021 地下水质分析方法 第15部分:总硬度的测定 乙二胺四乙酸二钠滴定法
- 预防艾滋病梅毒乙肝母婴传播干预措施
- 心理体检收费目录
- 雅鲁藏布江米林-加查段沿线暴雨泥石流危险度评价的中期报告
- 抗生素的正确使用与合理配比
- 读书分享读书交流会《局外人》课件
- 第十六章-常见骨关节疾病评定技术-2肩周炎评定
评论
0/150
提交评论